ndimage measurement's median provides result for missing labels

[jakirkham]

It appears that in some cases SciPy ndimage measurement's median function returns a result (other than nan) when a label doesn't exist. A reproducer of this odd behavior is provided below along with an environment file for use with conda.

Code snippet:

In [1]: import numpy as np

In [2]: import scipy.ndimage as spnd

In [3]: np.random.seed(0)

In [4]: a = np.random.random((4, 4))

In [5]: a
Out[5]: 
array([[ 0.5488135 ,  0.71518937,  0.60276338,  0.54488318],
       [ 0.4236548 ,  0.64589411,  0.43758721,  0.891773  ],
       [ 0.96366276,  0.38344152,  0.79172504,  0.52889492],
       [ 0.56804456,  0.92559664,  0.07103606,  0.0871293 ]])

In [6]: labels, labels_nb = spnd.label(a < 0.5)

In [7]: labels
Out[7]: 
array([[0, 0, 0, 0],
       [1, 0, 2, 0],
       [0, 3, 0, 0],
       [0, 0, 4, 4]], dtype=int32)

In [8]: np.unique(labels)
Out[8]: array([0, 1, 2, 3, 4], dtype=int32)

In [9]: spnd.median(a, labels, [1, 100])
Out[9]: array([ 0.4236548 ,  0.52889492])

Environment:

name: test
channels:
- conda-forge
- defaults
dependencies:
- appnope=0.1.0=py27_0
- backports.shutil_get_terminal_size=1.0.0=py27_1
- blas=1.1=openblas
- ca-certificates=2017.7.27.1=0
- certifi=2017.7.27.1=py27_0
- decorator=4.1.2=py27_0
- enum34=1.1.6=py27_1
- ipython=5.4.1=py27_0
- ipython_genutils=0.2.0=py27_0
- libgfortran=3.0.0=0
- ncurses=5.9=10
- numpy=1.13.1=py27_blas_openblas_200
- openblas=0.2.19=2
- openssl=1.0.2l=0
- pathlib2=2.3.0=py27_0
- pexpect=4.2.1=py27_0
- pickleshare=0.7.3=py27_0
- pip=9.0.1=py27_0
- prompt_toolkit=1.0.14=py27_0
- ptyprocess=0.5.2=py27_0
- pygments=2.2.0=py27_0
- python=2.7.13=1
- readline=6.2=0
- scandir=1.5=py27_1
- scipy=0.19.1=py27_blas_openblas_201
- setuptools=36.2.2=py27_0
- simplegeneric=0.8.1=py27_0
- six=1.10.0=py27_1
- sqlite=3.13.0=1
- tk=8.5.19=2
- traitlets=4.3.2=py27_0
- wcwidth=0.1.7=py27_0
- wheel=0.29.0=py27_0
- zlib=1.2.11=0
- pip:
  - backports.shutil-get-terminal-size==1.0.0
  - ipython-genutils==0.2.0
  - prompt-toolkit==1.0.14

[vbasu]

I looked at the implementation and it appears that any index that's not a valid label (not between 0 and the max label, not a positive value, not an integer value) will simply return the smallest value with label 0. Perhaps there is a reason for this, I don't know. I don't think it would be too difficult to alter this behavior, but I'm not sure what would be appropriate. Should it return nan? Should there be a warning?

[WarrenWeckesser]

If both labels and index are given, and there is a value in index that is not in labels, I think an error should be raised. What it does now makes no sense, so I'm calling it a bug.